Content Based Video Retrieval Using Spatiotemporal Salient Objects

In this paper, we propose a spatiotemporal salient objects-based approach for video retrieval. The spatiotemporal salient object is defined as the region sequence which is spatial salient and temporal continuous at the same time. As attention analysis is an effective mechanism for salient information selection, it provides a practical approach to narrow the semantic gap. Most existing methods extract salient region from static image, while little work has been done on the temporal consistency of salient region in dynamic image sequences, which is very important for content based video retrieval. We propose a fusion method for motion and spatial saliency integration to detect spatiotemporal salient objects, based on both attention analysis and interest point trajectory tracking. A novel description of trajectories' spatial context is also presented in this paper. Experiment results demonstrate the advantages of the proposed method in video content analysis and retrieval.